Google DiffusionGemma AI solving a complex puzzle, showcasing its fine-tuned capabilities and high success rate.

Editorial illustration for Google's DiffusionGemma Solves 85% of Puzzles After Minimal Fine-Tuning

Google's DiffusionGemma Solves 85% of Puzzles

By AI Daily Post Edited by Brian Petersen, Editor-in-Chief

August 9, 2026 • 4 min read

Google DeepMind published a technical report this week on DiffusionGemma, the diffusion-based language model it first released in mid-June. The headline number: after fine-tuning on less than ten percent of the original training token budget, the model solves 85 percent of a set of puzzle benchmarks, according to the report.

What makes that number notable is the starting point. DeepMind didn't build a new architecture from the ground up. It took Gemma-4-26B-A4B, an existing autoregressive model, and converted it into a diffusion model, the kind of system usually associated with image generation rather than text.

Instead of writing one token after another, DiffusionGemma refines blocks of 256 tokens at once, pulling coherent text out of noise the way Stable Diffusion pulls a picture out of static. On an Nvidia H100, DeepMind clocks it at roughly 1,500 tokens per second.

The report lays out a two-stage training process behind that jump in accuracy, and a tradeoff between speed and answer quality that shaped how the team built it. Before getting into the mechanics, here's how DeepMind's researchers frame the core idea themselves.

Building a new model from scratch wasn't necessary. The team started with the existing Gemma-4-26B-A4B and converted it into a diffusion model using less than ten percent of the original training token budget, according to the report.

Google's DiffusionGemma proves you don't need to train from scratch to build a text diffusion model - THE DECODER

Why this matters

DiffusionGemma is a retrofit, not a new architecture, and that's the part worth sitting with. Google DeepMind took Gemma 4, applied minimal fine-tuning, and got a model that solves close to 85 percent of a puzzle task where the base model scored zero. That gap tells us less about diffusion's raw power and more about how much latent capability was already sitting in Gemma, waiting for a different generation strategy to expose it.

For teams building on Gemma or similar open weights, that's a cheaper path than training from scratch, and it's worth testing on your own structured-output workloads, JSON generation, code repair, anything where the input already constrains most of the answer. The two-to-three refinement step number for those tasks is a real efficiency signal, not marketing. But we'd want to see how DiffusionGemma performs on open-ended generation, where there's no strong prior to lean on, before assuming this retrofit approach generalizes.

Watch for independent benchmarks outside puzzle-style tasks before betting production infrastructure on it.

Common Questions Answered

How much of the original training token budget did Google DeepMind use to convert Gemma-4-26B-A4B into DiffusionGemma?

Google DeepMind used less than ten percent of the original training token budget to convert the existing Gemma-4-26B-A4B autoregressive model into a diffusion-based model. This minimal fine-tuning approach proved highly efficient, demonstrating that building a new architecture from scratch was unnecessary to achieve significant performance improvements.

What puzzle benchmark performance did DiffusionGemma achieve after fine-tuning?

DiffusionGemma solved 85 percent of a set of puzzle benchmarks after minimal fine-tuning, according to Google DeepMind's technical report. This represents a dramatic improvement compared to the base Gemma model, which scored zero on the same puzzle task.

Why is DiffusionGemma being described as a retrofit rather than a new architecture?

DiffusionGemma is a retrofit because Google DeepMind took an existing autoregressive model (Gemma-4-26B-A4B) and converted it into a diffusion model rather than building a new architecture from the ground up. This approach demonstrates that significant latent capability was already present in the original model, waiting for a different generation strategy to expose it.

What does the performance gap between DiffusionGemma and the base Gemma model reveal about diffusion models?

The 85 percent puzzle-solving rate achieved by DiffusionGemma compared to zero percent for the base model reveals less about diffusion's raw power and more about how much untapped capability was already embedded in the original Gemma model. The gap demonstrates that applying a different generation strategy to an existing model can unlock substantial hidden performance gains.

Ship an AI product this weekend — no engineers required.

Structured, in-depth lessons on the exact no-code tools — not scattered tutorials.

The exact platforms, taught in depth
Build real, working projects
Our honest review + a reader discount

Read the review →